PetscErrorCode PCGAMGProlongator_Classical_Direct(PC pc, const Mat A, const Mat G, PetscCoarsenData *agg_lists,Mat *P)
{
PetscErrorCode ierr;
MPI_Comm comm;
PetscReal *Amax_pos,*Amax_neg;
Mat lA,gA; /* on and off diagonal matrices */
PetscInt fn; /* fine local blocked sizes */
PetscInt cn; /* coarse local blocked sizes */
PetscInt gn; /* size of the off-diagonal fine vector */
PetscInt fs,fe; /* fine (row) ownership range*/
PetscInt cs,ce; /* coarse (column) ownership range */
PetscInt i,j; /* indices! */
PetscBool iscoarse; /* flag for determining if a node is coarse */
PetscInt *lcid,*gcid; /* on and off-processor coarse unknown IDs */
PetscInt *lsparse,*gsparse; /* on and off-processor sparsity patterns for prolongator */
PetscScalar pij;
const PetscScalar *rval;
const PetscInt *rcol;
PetscScalar g_pos,g_neg,a_pos,a_neg,diag,invdiag,alpha,beta;
Vec F; /* vec of coarse size */
Vec C; /* vec of fine size */
Vec gF; /* vec of off-diagonal fine size */
MatType mtype;
PetscInt c_indx;
PetscScalar c_scalar;
PetscInt ncols,col;
PetscInt row_f,row_c;
PetscInt cmax=0,idx;
PetscScalar *pvals;
PetscInt *pcols;
PC_MG *mg = (PC_MG*)pc->data;
PC_GAMG *gamg = (PC_GAMG*)mg->innerctx;
PetscFunctionBegin;
comm = ((PetscObject)pc)->comm;
ierr = MatGetOwnershipRange(A,&fs,&fe); CHKERRQ(ierr);
fn = (fe - fs);
ierr = MatGetVecs(A,&F,NULL);CHKERRQ(ierr);
/* get the number of local unknowns and the indices of the local unknowns */
ierr = PetscMalloc(sizeof(PetscInt)*fn,&lsparse);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscInt)*fn,&gsparse);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscInt)*fn,&lcid);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscReal)*fn,&Amax_pos);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscReal)*fn,&Amax_neg);CHKERRQ(ierr);
/* count the number of coarse unknowns */
cn = 0;
for (i=0;i<fn;i++) {
/* filter out singletons */
ierr = PetscCDEmptyAt(agg_lists,i,&iscoarse); CHKERRQ(ierr);
lcid[i] = -1;
if (!iscoarse) {
cn++;
}
}
/* create the coarse vector */
ierr = VecCreateMPI(comm,cn,PETSC_DECIDE,&C);CHKERRQ(ierr);
ierr = VecGetOwnershipRange(C,&cs,&ce);CHKERRQ(ierr);
/* construct a global vector indicating the global indices of the coarse unknowns */
cn = 0;
for (i=0;i<fn;i++) {
ierr = PetscCDEmptyAt(agg_lists,i,&iscoarse); CHKERRQ(ierr);
if (!iscoarse) {
lcid[i] = cs+cn;
cn++;
} else {
lcid[i] = -1;
}
*((PetscInt *)&c_scalar) = lcid[i];
c_indx = fs+i;
ierr = VecSetValues(F,1,&c_indx,&c_scalar,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = VecAssemblyBegin(F);CHKERRQ(ierr);
ierr = VecAssemblyEnd(F);CHKERRQ(ierr);
/* determine the biggest off-diagonal entries in each row */
for (i=fs;i<fe;i++) {
Amax_pos[i-fs] = 0.;
Amax_neg[i-fs] = 0.;
ierr = MatGetRow(A,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
for(j=0;j<ncols;j++){
if ((PetscRealPart(-rval[j]) > Amax_neg[i-fs]) && i != rcol[j]) Amax_neg[i-fs] = PetscAbsScalar(rval[j]);
if ((PetscRealPart(rval[j]) > Amax_pos[i-fs]) && i != rcol[j]) Amax_pos[i-fs] = PetscAbsScalar(rval[j]);
}
if (ncols > cmax) cmax = ncols;
ierr = MatRestoreRow(A,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
}
ierr = PetscMalloc(sizeof(PetscInt)*cmax,&pcols);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscScalar)*cmax,&pvals);CHKERRQ(ierr);
/* split the operator into two */
ierr = PCGAMGClassicalGraphSplitting_Private(A,&lA,&gA);CHKERRQ(ierr);
/* scatter to the ghost vector */
ierr = PCGAMGClassicalCreateGhostVector_Private(A,&gF,NULL);CHKERRQ(ierr);
ierr = PCGAMGClassicalGhost_Private(A,F,gF);CHKERRQ(ierr);
if (gA) {
ierr = VecGetSize(gF,&gn);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscInt)*gn,&gcid);CHKERRQ(ierr);
for (i=0;i<gn;i++) {
ierr = VecGetValues(gF,1,&i,&c_scalar);CHKERRQ(ierr);
gcid[i] = *((PetscInt *)&c_scalar);
}
}
ierr = VecDestroy(&F);CHKERRQ(ierr);
ierr = VecDestroy(&gF);CHKERRQ(ierr);
ierr = VecDestroy(&C);CHKERRQ(ierr);
/* count the on and off processor sparsity patterns for the prolongator */
for (i=0;i<fn;i++) {
/* on */
lsparse[i] = 0;
gsparse[i] = 0;
if (lcid[i] >= 0) {
lsparse[i] = 1;
gsparse[i] = 0;
} else {
ierr = MatGetRow(lA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
for (j = 0;j < ncols;j++) {
col = rcol[j];
if (lcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold*Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold*Amax_neg[i])) {
lsparse[i] += 1;
}
}
ierr = MatRestoreRow(lA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
/* off */
if (gA) {
ierr = MatGetRow(gA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
for (j = 0; j < ncols; j++) {
col = rcol[j];
if (gcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold*Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold*Amax_neg[i])) {
gsparse[i] += 1;
}
}
ierr = MatRestoreRow(gA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
}
}
}
/* preallocate and create the prolongator */
ierr = MatCreate(comm,P); CHKERRQ(ierr);
ierr = MatGetType(G,&mtype);CHKERRQ(ierr);
ierr = MatSetType(*P,mtype);CHKERRQ(ierr);
ierr = MatSetSizes(*P,fn,cn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(*P,0,lsparse,0,gsparse);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(*P,0,lsparse);CHKERRQ(ierr);
/* loop over local fine nodes -- get the diagonal, the sum of positive and negative strong and weak weights, and set up the row */
for (i = 0;i < fn;i++) {
/* determine on or off */
row_f = i + fs;
row_c = lcid[i];
if (row_c >= 0) {
pij = 1.;
ierr = MatSetValues(*P,1,&row_f,1,&row_c,&pij,INSERT_VALUES);CHKERRQ(ierr);
} else {
g_pos = 0.;
g_neg = 0.;
a_pos = 0.;
a_neg = 0.;
diag = 0.;
/* local connections */
ierr = MatGetRow(lA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
for (j = 0; j < ncols; j++) {
col = rcol[j];
if (lcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold*Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold*Amax_neg[i])) {
if (PetscRealPart(rval[j]) > 0.) {
g_pos += rval[j];
} else {
g_neg += rval[j];
}
}
if (col != i) {
if (PetscRealPart(rval[j]) > 0.) {
a_pos += rval[j];
} else {
a_neg += rval[j];
}
} else {
diag = rval[j];
}
}
ierr = MatRestoreRow(lA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
/* ghosted connections */
if (gA) {
ierr = MatGetRow(gA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
for (j = 0; j < ncols; j++) {
col = rcol[j];
if (gcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold*Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold*Amax_neg[i])) {
if (PetscRealPart(rval[j]) > 0.) {
g_pos += rval[j];
} else {
g_neg += rval[j];
}
}
if (PetscRealPart(rval[j]) > 0.) {
a_pos += rval[j];
} else {
a_neg += rval[j];
}
}
ierr = MatRestoreRow(gA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
}
if (g_neg == 0.) {
alpha = 0.;
} else {
alpha = -a_neg/g_neg;
}
if (g_pos == 0.) {
diag += a_pos;
beta = 0.;
} else {
beta = -a_pos/g_pos;
}
if (diag == 0.) {
invdiag = 0.;
} else invdiag = 1. / diag;
/* on */
ierr = MatGetRow(lA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
idx = 0;
for (j = 0;j < ncols;j++) {
col = rcol[j];
if (lcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold*Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold*Amax_neg[i])) {
row_f = i + fs;
row_c = lcid[col];
/* set the values for on-processor ones */
if (PetscRealPart(rval[j]) < 0.) {
pij = rval[j]*alpha*invdiag;
} else {
pij = rval[j]*beta*invdiag;
}
if (PetscAbsScalar(pij) != 0.) {
pvals[idx] = pij;
pcols[idx] = row_c;
idx++;
}
}
}
ierr = MatRestoreRow(lA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
/* off */
if (gA) {
ierr = MatGetRow(gA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
for (j = 0; j < ncols; j++) {
col = rcol[j];
if (gcid[col] >= 0 && (PetscRealPart(rval[j]) > gamg->threshold*Amax_pos[i] || PetscRealPart(-rval[j]) > gamg->threshold*Amax_neg[i])) {
row_f = i + fs;
row_c = gcid[col];
/* set the values for on-processor ones */
if (PetscRealPart(rval[j]) < 0.) {
pij = rval[j]*alpha*invdiag;
} else {
pij = rval[j]*beta*invdiag;
}
if (PetscAbsScalar(pij) != 0.) {
pvals[idx] = pij;
pcols[idx] = row_c;
idx++;
}
}
}
ierr = MatRestoreRow(gA,i,&ncols,&rcol,&rval);CHKERRQ(ierr);
}
ierr = MatSetValues(*P,1,&row_f,idx,pcols,pvals,INSERT_VALUES);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(*P, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(*P, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscFree(lsparse);CHKERRQ(ierr);
ierr = PetscFree(gsparse);CHKERRQ(ierr);
ierr = PetscFree(pcols);CHKERRQ(ierr);
ierr = PetscFree(pvals);CHKERRQ(ierr);
ierr = PetscFree(Amax_pos);CHKERRQ(ierr);
ierr = PetscFree(Amax_neg);CHKERRQ(ierr);
ierr = PetscFree(lcid);CHKERRQ(ierr);
if (gA) {ierr = PetscFree(gcid);CHKERRQ(ierr);}
PetscFunctionReturn(0);
}
PetscErrorCode PCGAMGGraph_Classical(PC pc,const Mat A,Mat *G)
{
PetscInt s,f,idx;
PetscInt r,c,ncols;
const PetscInt *rcol;
const PetscScalar *rval;
PetscInt *gcol;
PetscScalar *gval;
PetscReal rmax;
PetscInt ncolstotal,cmax = 0;
PC_MG *mg;
PC_GAMG *gamg;
PetscErrorCode ierr;
PetscInt *gsparse,*lsparse;
PetscScalar *Amax;
Mat lA,gA;
MatType mtype;
PetscFunctionBegin;
mg = (PC_MG *)pc->data;
gamg = (PC_GAMG *)mg->innerctx;
ierr = MatGetOwnershipRange(A,&s,&f);CHKERRQ(ierr);
ierr = PCGAMGClassicalGraphSplitting_Private(A,&lA,&gA);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscInt)*(f - s),&lsparse);CHKERRQ(ierr);
if (gA) {ierr = PetscMalloc(sizeof(PetscInt)*(f - s),&gsparse);CHKERRQ(ierr);}
else {
gsparse = NULL;
}
ierr = PetscMalloc(sizeof(PetscScalar)*(f - s),&Amax);CHKERRQ(ierr);
for (r = 0;r < f-s;r++) {
lsparse[r] = 0;
if (gsparse) gsparse[r] = 0;
}
for (r = 0;r < f-s;r++) {
/* determine the maximum off-diagonal in each row */
rmax = 0.;
ierr = MatGetRow(lA,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
ncolstotal = ncols;
for (c = 0; c < ncols; c++) {
if (PetscAbsScalar(rval[c]) > rmax && rcol[c] != r) {
rmax = PetscAbsScalar(rval[c]);
}
}
ierr = MatRestoreRow(lA,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
if (gA) {
ierr = MatGetRow(gA,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
ncolstotal += ncols;
for (c = 0; c < ncols; c++) {
if (PetscAbsScalar(rval[c]) > rmax) {
rmax = PetscAbsScalar(rval[c]);
}
}
ierr = MatRestoreRow(gA,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
}
Amax[r] = rmax;
if (ncolstotal > cmax) cmax = ncolstotal;
ierr = MatGetRow(lA,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
idx = 0;
/* create the local and global sparsity patterns */
for (c = 0; c < ncols; c++) {
if (PetscAbsScalar(rval[c]) > gamg->threshold*PetscRealPart(Amax[r])) {
idx++;
}
}
ierr = MatRestoreRow(lA,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
lsparse[r] = idx;
if (gA) {
idx = 0;
ierr = MatGetRow(gA,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
for (c = 0; c < ncols; c++) {
if (PetscAbsScalar(rval[c]) > gamg->threshold*PetscRealPart(Amax[r])) {
idx++;
}
}
ierr = MatRestoreRow(gA,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
gsparse[r] = idx;
}
}
ierr = PetscMalloc(sizeof(PetscScalar)*cmax,&gval);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscInt)*cmax,&gcol);CHKERRQ(ierr);
ierr = MatCreate(PetscObjectComm((PetscObject)A),G); CHKERRQ(ierr);
ierr = MatGetType(A,&mtype);CHKERRQ(ierr);
ierr = MatSetType(*G,mtype);CHKERRQ(ierr);
ierr = MatSetSizes(*G,f-s,f-s,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(*G,0,lsparse,0,gsparse);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(*G,0,lsparse);CHKERRQ(ierr);
for (r = s;r < f;r++) {
ierr = MatGetRow(A,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
idx = 0;
for (c = 0; c < ncols; c++) {
/* classical strength of connection */
if (PetscAbsScalar(rval[c]) > gamg->threshold*PetscRealPart(Amax[r-s])) {
gcol[idx] = rcol[c];
gval[idx] = rval[c];
idx++;
}
}
ierr = MatSetValues(*G,1,&r,idx,gcol,gval,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatRestoreRow(A,r,&ncols,&rcol,&rval);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(*G, MAT_FINAL_ASSEMBLY); CHKERRQ(ierr);
ierr = MatAssemblyEnd(*G, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscFree(gval);CHKERRQ(ierr);
ierr = PetscFree(gcol);CHKERRQ(ierr);
ierr = PetscFree(lsparse);CHKERRQ(ierr);
ierr = PetscFree(gsparse);CHKERRQ(ierr);
ierr = PetscFree(Amax);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
